[OIiveirra]

Background

Iceberg V3 defines GEOMETRY and GEOGRAPHY with a coordinate reference system (CRS); GEOGRAPHY also identifies its edge algorithm. Their Parquet representation carries WKB bytes and spatial annotations. Doris needs to preserve that identity through catalog mapping, query execution and writeback, while keeping its existing STRING-based spatial functions compatible.

What problem does this PR solve?

Issue Number: N/A
Related PR: #67907
PR Type: Feature

Problem Summary:

Users need to query and write Iceberg V3 spatial columns without losing their CRS, edge algorithm or original WKB. Add native spatial types, recursive Parquet read/write and typed spatial functions. This requires consistent type transport, write validation before coercion, and explicit rejection of operations that lack spatial implementations. About 65% of added lines are tests and generated results; existing STRING spatial functions retain their behavior.

Details

• Change map: 216 files, +22,246/-593: production +7,679/-561; tests +13,680/-31; generated regression output +886/-0; build +1/-1. User documentation is in the separate website PR. Read the production changes in these parts:

  Part	Added/deleted lines	Purpose and main entry points
  Type identity and binary storage	+1,436/-148	Preserve kind/CRS/algorithm through FE/BE descriptors and raw bytes through ColumnSpatial and the shared binary SerDe.
  Iceberg mapping and Parquet I/O	+919/-26	Map nested schemas, retain spatial annotations and validate supported writes in the connector and Parquet writer.
  Admission and value execution	+1,415/-66	Preserve identity and required-list NULL rules before coercion, and admit only implemented expression/aggregate paths; start with ConnectorSpatialWriteNormalizer and ConnectorSpatialWriteValidator.
  Typed GIS and WKB validation	+3,909/-321	Bind typed constructors/accessors, preserve WKB and implement spherical calculations; start with functions_geo.cpp, wkb_parse.cpp and geography_calculation.cpp.

• Scope: Read/write top-level spatial columns, ARRAY elements, ordinary-key MAP values and STRUCT fields through Parquet. Preserve all seven basic ISO WKB kinds in XY/XYZ/XYM/XYZM, NULL/EMPTY and original bytes. ST_AsBinary preserves Z/M; typed text access and calculations reject non-XY inputs. Geography measurements and relations support default OGC:CRS84/spherical nonempty XY Point/LineString/Polygon, including typed ST_Equals, ST_Overlaps and ST_DWithin. Unsupported shapes/metadata and unresolved conversion-precision cases fail explicitly.
  Spatial Variant values, ORC materialization, native spatial GROUP BY/DISTINCT/set-operation keys, window partition keys and unsupported aggregate combinations are rejected. Preserve UNION ALL, supported spatial value aggregates/windows and ordinary-key/spatial-value operations. Paimon/Lance/GeoArrow, internal spatial storage, indexes, pruning, predicate pushdown, CRS conversion, typed WKT input and additional Geography algorithms remain outside scope.

• Design: Keep kind/CRS/algorithm in typed metadata because a generic binary or STRING carrier cannot identify them. Raw WKB is canonical so storage round trips retain the input bytes. Binary and spatial Parquet consumers share byte transfer; spatial validation stays in the spatial SerDe. FE write normalization retains ordinary keys/siblings and requiredness before coercion. Spherical predicates use original-coordinate checks to reject unresolved S2 boundary-rounding ambiguity. Scope decisions are recorded in this PR's review discussion; no separate design issue is linked.

• Usage: The typed constructor accepts hexadecimal WKB. For example, create a Point at (1, 2):

  SELECT ST_AsText(ST_GeometryFromWKBTyped(
      '0101000000000000000000F03F0000000000000040'));

  For an imported Iceberg spatial column, use ST_AsText(geom) or hex(ST_AsBinary(geom)) to inspect it; matching target schemas support INSERT SELECT while preserving spatial identity.

• Default behavior: No new runtime flag is introduced. Eligible Iceberg fields map to native spatial types automatically. Spatial writes require Iceberg V3 and Parquet, with an explicit connector capability for nested leaves. Nondefault CRS/algorithm metadata is retained without conversion; property-referenced projjson: writes are rejected before file creation.

• Compatibility: Participating FEs/BEs must support the new types; mixed 4.x/5.0 execution and downgrade support for spatial metadata are outside scope. Connector plugins must match API 31.0, distinct from upstream API 30.0. Existing VARBINARY Block layout and STRING constructors remain unchanged. Typed MySQL values use binary metadata; generic Flight clients need text/binary projections because direct spatial Flight schemas are unsupported. Retain the BE constant-folding exclusion because FE cannot decode spatial PValues.

• Behavior change: Iceberg spatial fields are exposed as native types with binary MySQL results. Connector plugins must be rebuilt for API 31.0, including plugins that only expose ordinary columns.

• Persistence and failure handling: Preserve spatial parameters in schema/type transport and Parquet annotations. Validate spatial identity and supported schemas before coercion or file creation. Reuse Iceberg's existing commit/conflict and cleanup lifecycle; DDL preflight checks SDK-observed schemas, without claiming a server-wide invariant for REST rebases the client cannot observe.

• Follow-up plan: Nullable V2 batching and general dense-Polygon optimization remain deferred.

Testing

• Local validation at d89b8a795bc: 337 FE tests passed, including the complete Connector SPI suite and selected Iceberg mapping/DDL/write-plan tests.
• Local validation at e993d2e9d03 and c3c75caa110: 45 binary/spatial SerDe BE ASAN tests and 74 FE validator/normalizer/converter tests passed. These cover explicit ORC rejection and nested required-list NULL handling.
• Earlier recorded integration validation at 8a774c249e29: 360 FE tests, 132 Geography BE tests, and the GIS/Iceberg suites passed locally, with 70 GIS and 131 Iceberg result sections. These results predate subsequent changes.
• On current head e27809effd8, selected BE tests did not execute: local Lance/Rust dependency setup blocked the build. The new BSON-isolation regression and latest required-list external SQL cases remain unexecuted locally. The external suite requires enableIcebergTest and working Spark/MinIO fixtures. Latest-head full community CI has not been verified here; earlier passes are not a current-head CI verdict.

Notes for Reviewers

• The latest merge keeps BSON decoding on concrete VARBINARY SerDe only; spatial columns retain the default BSON rejection. It adds no spatial MongoDB support.
• Detailed precision/error cases are retained in the Geography tests and the spatial regression suites. Unsupported operations are deliberate boundaries, not promises of bytewise spatial SQL equality.

Release note

Support native Iceberg V3 GEOMETRY/GEOGRAPHY schema mapping and Parquet read/write, including ARRAY elements, ordinary-key MAP values and STRUCT fields. Preserve CRS, edge algorithm and raw ISO WKB including Z/M, with explicit XY calculation limits and unsupported-operation errors. Connector plugins must match API 31.0.

Check List (For Author)

• Behavior changed: Yes (see Details)
• Does this need documentation: Yes (apache/doris-website#4239)
• Pre-submission status: Partially verified: latest-head BE unit execution is blocked by local dependencies; the new required-list external cases and latest full community CI have not been verified here.

AI Disclosure

• Tool: Codex (model: not stated)
• Generated: Assisted with implementation, tests, review-driven fixes and this description.
• Human review: not stated

[OIiveirra]

/review

[github-actions]

Static review of e27809effd8a021885afb0d00be0e99d0fc12a14, covering all 216 changed files and their applicable upstream/downstream paths. Two new P2 findings: the connector version-wiring regression expects API 30 although this PR serves API 31, and ordinary Polygon/Line containment performs a full polygon-edge scan for every line vertex. See the inline comments. No additional user focus was supplied.

Critical checkpoint conclusions:

Checkpoint	Conclusion
Goal and evidence	Native Iceberg V3 spatial identity/read/write plumbing matches the documented subset; unit and regression cases cover positive and rejected operations. The version test has the verified mismatch reported inline.
Scope and reuse	Broad but cohesive feature change; shared binary SerDe/serialization and existing planner/connector paths are reused. No unrelated production change identified.
Concurrency	External schema/format changes can race preflight. The DDL adapter checks actual SDK commit candidates and preserves CAS/retry behavior. No new worker/shared-state race verified.
Lock cost and ordering	No new lock hierarchy or heavy work under an introduced lock found.
Lifecycle and initialization	Owning spatial copies, arena retention/trim, non-owning DDL adapters and writer-close-before-cleanup paths were traced. No new supported-path lifetime or static-initialization defect verified.
Configuration	No new runtime configuration item.
Compatibility	Connector API 31 requires plugin rebuilds. VARBINARY Block layout and legacy STRING GIS routes are retained. Mixed 4.x/5.0 spatial execution remains explicitly excluded by the approved scope, not fixed by this PR.
Thrift/protobuf allocation	Read every touched definition on HEAD and live-verified 3.1/4.0/4.1/4.2/master refs. New fields are optional; explicit spatial IDs avoid the reserved FILE identities. No introduced collision/reuse found.
Parallel paths	Checked scalar/nested types, typed/legacy functions, INSERT/UPDATE/MERGE, schema/default/equality-delete transport, aggregate/state/rewrite admission, grouping/set/window keys, runtime filters and client projections. No additional verified omission.
Conditional/error paths	Unsupported dimensions/metadata/operations fail explicitly. NULL provenance and cast checks preserve ordinary siblings and original errors. Dynamic required LIST NULLs reach Arrow's explicit required-leaf rejection.
Test coverage	Inspected positive/negative cases for metadata, nullable/const materialization, DDL retries, nested requiredness, planner identities, ownership and numerical boundaries. No build or test executed for this review.
Changed results	Inspected regression projection shapes, NULL/EMPTY/byte expectations and boundary controls. No further substantiated mismatch; exhaustive numerical recomputation was not performed.
Observability	Existing scan/connector profiling and explicit operation/metadata errors cover the new paths; no additional logging/metric defect verified.
Persistence and failover	Spatial parameters use existing Gson/Thrift/protobuf metadata channels; no new EditLog operation. External snapshot commit/retry remains delegated to Iceberg. No new failover defect verified.
Data writes and atomicity	Schema/write admission and failed-writer cleanup were traced through commit reporting. No new supported-path atomicity or cleanup defect verified.
FE/BE transport	Kind/CRS/algorithm are retained across descriptors, nested schema carriers and serialization. Remote constant folding still excludes spatial values FE cannot decode.
Performance and final rethink	Verified a distinct quadratic Polygon/Line side-admission path: a 320,070-byte input pair forces 100,020,000 point-edge distance calls before containment. Reported inline; counts derive from source, not a benchmark. Other ownership/resolution/candidate paths were rechecked.

Prior P1 dispositions were independently checked: nullable COUNT (4269360799), FILE ID reuse (4384986088), portable wide arithmetic (4402847165), dimension admission (4434477659) and ORC casts (4469669669) are fixed. Bearing (4409568560) uses the full-angle sine; identity partitions (4426807465) are rejected by the pinned Iceberg transform; polygon interiors (4428917878) are included by the pinned S2 query and target defaults. Mixed-version finding 4430686953 is outside the explicit maintainer-approved scope. No existing in-scope P0/P1 remains independently confirmed, so the blocking-ID set is empty. Known prior findings were not duplicated.

Validation limits: static inspection only; no current-head build, test, benchmark, external-engine run or exhaustive numerical proof. The false/true scanner-flag loops both select current V2 and are not independent V1/V2 validation. Earlier author-reported passes do not establish current-head CI. Documented unsupported formats/algorithms, direct spatial Flight schemas and mixed-version execution remain scope boundaries. No unresolved concrete review candidate remains.

Review converged after 3 round(s).

[OIiveirra]

run buildall

[OIiveirra]

/review

[github-actions]

Reviewed exact head 5cdac8ce51a404438b08cc100591851330b5fb97. Three new P2 findings are attached: inconsistent UNION ALL branch types after spatial NULL normalization; incorrect MAP-constructor exception expectations in the external suite; and premature spatial validation for existing Iceberg BINARY fields. Each was independently verified and deduplicated against the existing threads.

The coverage inventories account for all 231 changed paths: FE/regression 149 and BE/gensrc 82, with a separate risk review of Geography precision and NULL normalization. There was no additional user-provided focus.

Critical checkpoint conclusions:

1. Goal and behavior: Native Iceberg V3 spatial metadata, raw WKB storage, Parquet I/O and constrained typed functions are implemented. The attached supported-write/read failures prevent a clean correctness conclusion.
2. Scope and reuse: This is a broad spatial feature across type, connector and execution boundaries. Shared binary serialization, SerDe and existing planner/connector helpers keep the changes related; no unrelated feature change was identified.
3. Concurrency and locks: No new thread or lock protocol was introduced. Existing SDK CAS/retry and asynchronous file-close interactions were traced; no additional verified concurrency defect was found.
4. Lifecycle and initialization: Reviewed column COW/arena ownership, owning aggregate/Field states, nullable/constant paths and failed-open cleanup before object deletion. No additional verified lifecycle or static-initialization issue was found.
5. Configuration: No new runtime configuration needs dynamic propagation. Annotation interpretation is a reader contract, and the metadata-cache key includes that mode.
6. Compatibility: Connector API 31.0 and the API-30 rejection test agree. Legacy STRING spatial signatures remain separate. The BINARY read regression is attached. Mixed 4.x/5.0 spatial execution remains the existing explicit maintainer scope decision; participating nodes must support the new types.
7. Thrift/protobuf: Compared every touched definition with live-verified master, branch-3.1, 4.0, 4.1 and 4.2. New fields are optional; spatial enum allocation preserves branch-4.1 FILE's protobuf45/Thrift47 identities. No new collision was found.
8. Parallel paths: Covered INSERT VALUES/SELECT, UPDATE, MERGE/changelog/rewrite; CREATE/ALTER; scalar/nested/state/window/aggregate admission; generic Hive/load versus Iceberg; and data/default/equality-delete carriers. The UNION and Iceberg BINARY gaps are attached.
9. Conditional checks: Guards correspond to concrete unsupported operations, formats, metadata or NULL provenance. No additional verified guard defect was found after checking upstream guarantees.
10. Test coverage: Substantial positive/negative FE, BE and integration coverage exists in source. Missing combinations include whole-NULL/non-NULL UNION containers with ordinary sibling widening and actual-file decoding through an Iceberg BINARY schema.
11. Expected results: Current golden section shapes, including the corrected three-column Z/M projections, were inspected. The MAP-constructor expectation mismatch is attached. Generated-output provenance and current-head execution were not verified.
12. Observability: Existing rejection paths retain relevant row/type/column/CRS/format context. No additional verified logging or metrics defect was identified.
13. Persistence and failover: Spatial identity uses existing recursive type/schema carriers; existing Iceberg commit/conflict handling is retained. No new journal or transaction state machine was introduced.
14. Writes and atomicity: Checked pre-coercion identity validation, format/CRS admission and failed-writer cleanup. Dynamic required-list NULLs are also rejected by the existing Arrow 24 writer boundary. No additional verified commit/atomicity defect was found.
15. FE/BE transport: Kind, CRS and algorithm propagate through scalar/column descriptors and historical schemas. The recursive BE-folding exclusion remains for values FE cannot decode; no new session-variable transport is needed.
16. Performance: Reviewed arena retention, binary transfer, spatial noding/index recall and indexed Polygon/Line containment. Known nullable-selection batching and general dense-Polygon work remain previously discussed limits. No benchmarks or measured speed claims were made.
17. Other issues and existing reviews: All three accepted new findings are inline. Prior P1 COUNT-nullability, FILE-ID reuse, portable-wide-arithmetic, dimensional-validation and ORC-cast concerns are fixed in source. The bearing allegation is not the actual code; the polygon-distance and spatial-partition allegations are ruled out by the pinned S2 query and Iceberg identity transform. No existing in-scope P0/P1 finding was independently confirmed as still applicable, so none is duplicated or carried as a blocking ID.

Validation was static only: no builds, SQL execution, tests, benchmarks, fixture runs, source edits or submodule initialization. Runtime behavior and exhaustive numerical correctness are not claimed. There are no unresolved concrete candidates or unassigned changed paths after the final challenge; process convergence is recorded separately by the submission helper.

Review converged after 2 round(s).

[OIiveirra]

/review

[github-actions]

Reviewed head b58b2e216984bfee7ae1c7e4f0ffafcff31e6df3 statically. One new P2 finding concerns opaque BINARY equality-delete inputs; see the inline comment. No outstanding in-scope existing P0/P1 finding was independently confirmed.

Critical checkpoint conclusions:

• Goal and correctness: traced Iceberg schema → FE spatial identity/admission → Thrift → BE WKB columns/Parquet → GIS or MySQL/Iceberg output. The intended read/write feature is implemented broadly; the separate equality-delete consumer still has the inline defect.
• Scope and clarity: the change is broad because it introduces types across the full pipeline. Shared binary serialization/SerDe avoids duplicating the carrier, although NULL normalization and raw/S2 certification remain complex.
• Concurrency and locks: no new production thread, lock order or mutable global lifecycle was introduced in the reviewed paths. DDL validation retains the SDK's commit/CAS mechanism; test counters are thread-local.
• Lifecycle/static initialization: column/state ownership, arena trimming, split-local schema mutation and failed-writer close-before-delete were inspected. No additional ownership or cross-translation-unit initialization defect was verified.
• Configuration: no new runtime setting requires dynamic propagation.
• Compatibility: VARBINARY's block layout is retained; connector API 31 rejects pre-spatial plugins. Spatial participants must support these types; mixed 4.x/5.0 spatial execution/downgrade is the existing explicit scope decision.
• Thrift/protobuf: compared every touched definition against live master and branch-3.1/4.0/4.1/4.2 refs. New fields are optional and new enum values avoid the release FILE identities. No new collision was found.
• Parallel paths: checked legacy STRING versus typed GIS, annotated/unannotated and explicit BINARY files, nested containers, hidden/deleted keys, NULL/const, aggregate/state/window and DML paths. Standalone tombstone decoding needs the same table-aware BINARY handling as data decoding. The concrete Parquet downcast also conflicts with the documented neutral reader/mapping boundary; place shared materialization policy at that boundary.
• Conditional checks/errors: unsupported dimensions, metadata, shapes and operation families fail explicitly; typed storage and ST_AsBinary retain raw bytes. No additional reachable silent-continuation issue was verified.
• Test coverage: inspected added BE/FE tests and regression cases for identity, required lists, coercion, NULLs, byte round trips, GIS precision and failure cleanup. Opaque annotated BINARY tombstones/hidden keys lack the coverage described inline.
• Expected results: checked the changed regression outputs against their projections and assertions; the previous MAP-error and Z/M projection mismatches are repaired. Numerical results were reviewed against source logic, without executing Doris.
• Observability: existing query statuses/profiles remain available; no new critical logging/metric omission was substantiated. Performance conclusions are source-based, with no benchmark claim.
• Persistence/failover: spatial parameters use the existing type transport/persistence mechanisms. Iceberg DDL validates SDK-observed commit candidates and retains existing conflict handling; unseen REST server rebases remain the documented limit.
• Writes/atomicity: existing Iceberg commit/rollback machinery is reused. Inspected the output-file failure/unwind changes and required-list writer rejection; no separate new atomicity defect was established.
• FE–BE propagation: kind, CRS and algorithm are carried recursively, MySQL metadata is binary, and spatial remote constant folding remains excluded because FE cannot decode those values.
• Performance: reviewed column copying/erasure, reusable materialization and GIS work-count guards. Nullable V2 batching and general dense-Polygon/circle limits are already discussed follow-ups and are not duplicated.
• Other concerns/user focus: no additional user focus was supplied. The existing UNION NULL-provenance concern remains reachable under a direct LIMIT/SORT wrapper; it is deduplicated, not posted again. The latest branch common-type repair is present. Other confirmed repaired/disproved P1 threads are not carried as blockers.

Review convergence is assessed separately from feature correctness and runtime validation. This review ran no build, test, benchmark, cluster, source edit or submodule initialization; author-reported runs are not treated as execution by this review.

Review converged after 2 round(s).

[OIiveirra]

/review

[github-actions]

Codex automated review failed and did not complete.

Error: Step 'Resolve review inputs' ended with failure; inspect its logs for the original error.
Workflow run: https://github.com/apache/doris/actions/runs/43238469865

Please inspect the workflow logs and rerun the review after the underlying issue is resolved.

[OIiveirra]

/review

[github-actions]

Static review of PR 69032 at exact head b58b2e2.

Verified finding:

• F1 [P1] MapAggregateFunction's new legality check rejects only direct spatial MAP keys. Nested keys such as ARRAY, MAP<INT,GEOGRAPHY>, and STRUCTshape:GEOGRAPHY can reach the BE map combinator, where dispatch_switch_all has no complex-key case and DORIS_CHECK aborts instead of returning an analysis error. This affects count_map, sum_map, avg_map, min_map, and max_map. The finding is inline at the exact changed guard.

Review checkpoints:

• The feature goal is native Iceberg V3 GEOMETRY/GEOGRAPHY mapping, Parquet read/write, metadata preservation, and supported typed functions. The inspected FE, BE, v1/v2 Parquet, Iceberg, connector, and protocol paths implement that goal except for F1.
• The change is broad but internally separated into type/column, parser/calculation, connector validation, and Parquet/Iceberg layers. No new concurrency primitive was introduced; writer close and reader lifecycle paths were inspected.
• Spatial metadata is carried through FE/BE descriptors and Parquet schema conversion. Protocol IDs reserve branch-4.1 FILE identities and new metadata fields are optional. Existing P1 review finding 4430686953 remains applicable because mixed-version Thrift execution is explicitly outside this PR's compatibility scope.
• Parallel v1/v2 readers, raw BINARY overlays, nested LIST/MAP/STRUCT mappings, equality-delete handling, cache-key flags, and writer CRS preflight were checked. Existing annotated-BINARY/equality-delete discussions were treated as known context and not duplicated.
• Column ownership, arena compaction, nullable handling, WKB validation, error propagation, and aggregate/function dispatch were inspected. F1 is the remaining newly verified dispatch gap.
• Tests and generated outputs were inspected as static evidence only. No build, test, benchmark, or runtime validation was executed in this review.
• No additional user review focus was supplied. Existing current-head P2 discussions 4545596368 and 4551380707 remain known context and are not duplicated.

The review converged in round 1: all assigned specialist scopes completed, two returned NO_NEW_VALUABLE_FINDINGS, and the remaining candidate was independently verified and accepted.

Review converged after 1 round(s).

Existing P0/P1 findings confirmed for this head: #69032 (comment)

[OIiveirra]

/review

[github-actions]

One new P2 correctness finding: the spatial BINARY override can undo a timestamp sibling rewrite in nested Iceberg equality-delete files, causing an otherwise valid scan to fail. It is distinct from the earlier missing-BINARY-override thread. The inline comment contains the trigger and fix direction.

Static coverage: all 234 changed paths (149 FE/regression, 70 BE I/O/type/protocol, 15 GIS), complete changed test bodies and regression results, plus an independent risk review. No additional user-provided focus was supplied.

Critical checkpoint conclusions:

1. Goal and proof: native spatial identity/WKB, Iceberg v3 Parquet I/O and deliberately bounded Geography calculations are wired end to end; the reported delete-reader interaction remains incorrect. Relevant positive/negative tests are present and were inspected.
2. Scope and reuse: the large change follows one type feature through its consumers; shared binary serialization, validation and schema helpers reuse existing mechanisms.
3. Concurrency and locks: no new worker or lock order found. DDL candidate validation retains existing Iceberg CAS/retry behavior; reader overrides are split-local and separate from cached metadata.
4. Lifecycle: column/aggregate payloads own their bytes; window erasure reclaims dead arenas. Writer failure cleanup closes/destroys upload owners before deletion. No new cross-TU initialization dependency found.
5. Configuration: no new runtime setting; existing timestamp mapping is the reported trigger. SQL-mode rendering and internal byte-mode cache identity were checked.
6. Compatibility: legacy STRING GIS and ordinary VARBINARY behavior remain separate; the VARBINARY Block layout is unchanged. Connector API31 and its version gate/baseline agree. Mixed 4.x/5.0 spatial execution and downgrade remain explicitly unsupported under the existing maintainer scope decision; this review does not certify them.
7. Thrift/protobuf: all touched definitions compared with branch-3.1/4.0/4.1/4.2 and master refs. Spatial enums avoid FILE's existing IDs; new metadata fields are optional and no new collision was found.
8. Parallel consumers: checked typed/legacy, scalar/nested, plain/dictionary, NULL/constant, projected/hidden delete keys, constructors/aggregates/state/combinators, INSERT/UPDATE/MERGE/changelog, and data/delete-only writers. The combined BINARY/timestamp tombstone case is the verified omission.
9. Conditions: metadata/format/dimension/EMPTY restrictions and NULL-provenance checks have concrete paths and negative controls. Ordinary sibling casts, input scope, ExprIds, UNION common types and selected MERGE-branch evaluation were traced.
10. Tests: inspected unit, actual-file and external regression coverage. Missing combined nested BINARY/timestamptz delete-key coverage is requested inline.
11. Results: changed golden rows match query/fixture shapes on inspection; no additional result mismatch found. Generation provenance and runtime numeric results were not verified.
12. Observability: unsupported operations/WKB failures retain analysis or Status context; cleanup logs identify files. No new per-row INFO logging or concrete additional metric defect found.
13. Persistence and failover: CRS/algorithm use existing type persistence and descriptor paths. No new EditLog, visible-version or MoW logic; Iceberg retains existing metadata commit ownership.
14. Writes and failure atomicity: full-schema admission and pre-file CRS validation were checked, along with existing commit/rollback and failed-open cleanup. No separate new partial-commit, leak or deadlock found.
15. FE/BE transport: scalar/column Thrift, protobuf and connector copies retain spatial parameters. Constant folding excludes unsupported spatial PValues; MySQL and typed ST_AsBinary use binary carriers.
16. Performance: reviewed owned filtering/erasure, partial TopN, raw-coordinate calculations and indexed admission. Already-reported nullable V2 batching and general dense-polygon/long-arc work remain known limits and are not reposted; no universal linear bound or benchmark result is claimed.
17. Final challenge and prior reviews: rechecked candidates, callers, alternate paths and anchors. Historical P1 COUNT, FILE IDs, Apple ARM arithmetic, dimensional validation, ORC and nested MAP keys are repaired; bearing, polygon-interior distance and identity-partition allegations are contradicted by current code/library contracts and tests. No in-scope existing P0/P1 remains independently confirmed; the mixed-version scope exception is stated above.

Validation was static inspection only. No builds, tests, benchmarks, cluster execution, source changes or submodule initialization were performed. Runtime/platform/ASAN/external interoperability validation remains unverified.

Review converged after 2 round(s).

[OIiveirra]

/review

[OIiveirra]

run buildall

[github-actions]

Review of PR #69032 at 4d7ce45ce69a02ab2ff547496eac9dfc505a7522: no new actionable finding was verified in the supported feature scope. No additional user review focus was supplied. Existing findings were not reposted.

Static review covered all 234 paths in the supplied aggregate diff: 82 BE, 143 FE, three shared protocol definitions and six regression SQL/output files. Frontend/backend complete reviews and a separate risk review examined related invocation chains and tests. No builds, tests, benchmarks, source edits or submodule initialization were performed. Author-reported test passes were treated as prior reports, not validation executed by this review.

Critical checkpoint conclusions:

• Goal and proof: spatial kind/CRS/algorithm and raw WKB are carried through Iceberg mapping, Parquet I/O and typed execution. Positive/negative tests were inspected; source inspection does not establish a current-head runtime pass.
• Scope and reuse: the broad change follows the necessary type, connector, planner, reader/writer and function boundaries. Binary payload transfer is shared while spatial validation remains separate. No additional scope/refactoring defect identified.
• Concurrency and locks: function scratch, normalization state and semantic reader trees are query/split owned; cached footer metadata stays immutable and cache keys include annotation interpretation. Iceberg DDL validates the SDK-observed commit candidate across retries. No new thread-entry, lock-order or heavy-under-lock defect identified.
• Lifecycle and initialization: column copies and aggregate states own WKB; mutable filtering preserves arena ownership and erasure reclaims discarded storage. Writer error cleanup closes pending uploads before deletion. Wide arithmetic has a portable fallback for narrow long double. No additional ownership/static-initialization defect identified.
• Configuration: no new runtime configuration or dynamic-update obligation.
• Compatibility and API/storage changes: existing VARBINARY block layout is retained; connector API is deliberately bumped to 31.0, with current31/prior30 probes aligned. Participating nodes must support spatial types. Mixed 4.x/5.0 spatial execution is an explicit maintainer-approved scope exclusion (decision, existing disposition); it is not reposted or counted as an in-scope P1.
• Thrift/protobuf IDs: complete touched definitions were compared with current live-matching branch-3.1, branch-4.0, branch-4.1 and branch-4.2 definitions. New metadata fields are optional; spatial enum allocations avoid branch-4.1 FILE's protobuf45/Thrift47 identities. No new collision or ID reuse found.
• Parallel paths: scalar/nested, plain/dictionary, nullable/constant, data/equality-delete, typed/opaque-BINARY and insert/update/merge paths were inspected. ORC spatial paths explicitly reject before invalid casts; Hive/load BINARY consumers retain opaque bytes. Sort/spill and position-delete writer entry points were also traced.
• Conditional and error behavior: unsupported metadata, dimensions, kernels and unresolved calculation ambiguity have explicit rejection paths. Actual NULL required container values are rejected recursively by VIcebergTableWriter::_write_prepared_block -> iceberg::validate_write_block before partition dispatch or Arrow conversion; optional parent NULLs correctly mask children. No reachable required-container write defect remains.
• Test coverage: inspected unit/regression additions cover transport, ownership, NULLs, complex schemas, dimensions, negative admission, BINARY overrides, equality deletes, numerical cases and work counts. Existing runtime write-validator tests also cover required elements/children and rejection before writer creation. Runtime/platform/cross-writer execution remains unverified here.
• Expected results: changed regression SQL and generated output sections were statically compared for shape, row counts and expectations, including corrected Z/M projections. No additional mismatch was substantiated. Current-head output generation and independently computed numerical truth for every spherical fixture were not verified by execution.
• Observability: errors retain type/row/file context at existing boundaries and reader profile paths remain intact; no new high-volume INFO logging defect identified. Work-count hooks are present, but logs, metrics and performance counters were not measured.
• Persistence and failover: spatial parameters use the existing catalog/schema/type serialization paths. Internal spatial storage is rejected; no internal rowset, visible-version, delete-bitmap or EditLog transaction protocol changes. Iceberg metadata commit/retry and cleanup retain their existing ownership.
• Writes and atomicity: data-file writes use the pinned target schema and existing commit/rollback lifecycle. Recursive target-schema value validation precedes all production partition-writer routes; sorted/spilled data has already passed that validation. The direct alternate Parquet-writer consumer is restricted to fixed position-delete fields. No new write-atomicity or value-loss defect identified.
• FE/BE transport: connector/catalog types, Thrift, PTypeDesc, PColumnMeta, MySQL binary metadata and recursive spatial constant-fold exclusion were inspected. Direct Flight spatial schemas remain unsupported as documented; text/binary projections are the supported client route.
• Performance and memory: in-place filtering, periodic arena compaction, byte-oriented WKB parsing and indexed/grouped noding address earlier redundant work. Query-owned CTE memoization avoids repeated producer expansion. Nullable-selection batching and general dense-polygon optimization remain known existing discussion limits; no new distinct performance defect or benchmark claim.
• Additional risks and prior findings: current timestamp-sibling/BINARY overrides synchronize split-local sibling types before rebuilding parents; UNION normalization reconciles branch types and preserves output identities; recursive spatial MAP-key rejection and API31 wiring are present. Earlier in-scope P1 reports are fixed or excluded by current source/dependency guards. No existing in-scope blocking comment ID remains to carry forward, and no concrete candidate or initial risk item remains unresolved.

The conclusion is bounded to static inspection of this head. Runtime/interoperability, exhaustive numerical-oracle and benchmark validation remain limits of this review, not claims of completed execution.

Review converged after 3 round(s).